熵
信息量
概率可能性的投影空间大小,单位比特。记随机离散变量的分布列与信息量I为:
[
x
0
x
1
.
.
.
x
n
p
0
p
1
.
.
.
p
n
]
I
(
x
i
)
=
−
l
o
g
(
p
(
x
i
)
)
\begin{alignedat}a &\begin{bmatrix} x_0 & x_1 & ... &x_n\\ p_0 & p_1 & ... &p_n \end{bmatrix}\\\\ &I(x_i) = -log(p(x_i)) \end{alignedat}
[x0p0x1p1......xnpn]I(xi)=−log(p(xi))
熵
熵(Entropy):一个系统所有事件的不确定性之和;
相对熵(Relative Entropy):同一个随机变量的两个不同分布间的差别描述;
交叉熵(Cross Entropy):使用分布
q
(
x
)
q(x)
q(x)表示目标分布
p
(
x
)
p(x)
p(x)的困难程度,以分布
q
q
q描述拟合分布
p
p
p。
H
(
x
)
=
E
(
L
e
n
b
i
t
)
=
E
(
l
o
g
(
1
p
i
)
)
=
−
∑
i
=
1
n
p
i
l
o
g
(
p
i
)
D
K
L
(
p
∣
∣
q
)
=
∑
i
=
1
n
p
i
l
o
g
(
p
i
q
i
)
=
∑
i
=
1
n
p
i
(
l
o
g
(
p
i
)
−
l
o
g
(
q
i
)
)
H
(
p
,
q
)
=
E
(
L
e
n
b
i
t
)
=
E
(
l
o
g
(
1
q
i
)
)
=
−
∑
i
p
i
l
o
g
(
q
i
)
\begin{alignedat}a H(x) &= E(Len_{bit}) = E(log({1\over p_i})) = - \sum_{i=1}^np_ilog(p_i)\\ D_{KL}(p||q) &= \sum_{i=1}^np_ilog({p_i \over q_i}) = \sum_{i=1}^np_i(log(p_i)-log(q_i))\\ H(p,q) &= E(Len_{bit}) = E(log({1\over q_i})) = -\sum_{i}p_ilog(q_i) \end{alignedat}
H(x)DKL(p∣∣q)H(p,q)=E(Lenbit)=E(log(pi1))=−i=1∑npilog(pi)=i=1∑npilog(qipi)=i=1∑npi(log(pi)−log(qi))=E(Lenbit)=E(log(qi1))=−i∑pilog(qi)